“OpenAI’s AgentKit bridges the gap between chatbots and true digital coworkers — here’s why developers are calling it a revolution.”

🔍 Key Updates & New Features

1. Agent Kit launch

   
   

   • OpenAI introduced an Agent Kit—a toolkit to help developers and enterprises build AI agents—from prototyping through production.

   • This suggests more packaged, end-to-end support (not just SDKs) for building, managing, deploying agents.

     
     

2. Realtime API & voice agents improvements

   
   

   • The Realtime API is now generally available, with support for:

     • Remote MCP servers

     • Image inputs

     • Phone calling via SIP (Session Initiation Protocol)These allow voice agents to access richer tools/context and to make external calls.

   • There’s a new RealtimeAgent feature (in TypeScript SDK) for voice agents, along with better tracing for real-time flows, and improvements to speech-to-speech.

     
     

3. TypeScript / JavaScript Agents SDK support

   
   

   • The Agents SDK, initially in Python, now also has a TypeScript / JS version, with feature parity: handoffs, guardrails, tracing, and other core agent primitives.

   • The versioning and release process for the JS SDK is documented (minor/patch version rules)

   • As of now, the npm package @openai/agents is live, with recent updates (version 0.1.0)

     
     

4. Agent durability / long-running workflows via Temporal integration

   
   

   • OpenAI partnered with Temporal to add Durable Execution to agents built with the Agents SDK, available in public preview. That means the agents can reliably resume, retry, and continue operations even under failures or network issues.

   • This helps handling rate limits, network flakiness, crashes etc. without complicating your agent code.

     
     

5. Tracing, observability, and instrumentation

   
   

   • Built-in tracing is a core feature: you can visualize, debug, and monitor agent runs, tool usage, handoffs etc.

   • Phoenix / OpenTelemetry instrumentation: agents now support better visibility into LLM calls, tool usage, spans etc.

   • Support for MCP Span Tool Info was added to capture span data for MCP tool listings.

     
     

6. Agent SDK enhancements & version changes

   
   

   • In version 0.2.0 of the Python Agents SDK: some types changed (e.g. some places that accepted Agent now accept AgentBase) — important if you are upgrading.

   • Recent patches (v0.2.9, v0.2.10, v0.2.11) fixed issues like streaming logic, tool call events, docs, preserving “thinking” blocks, connector fixes etc.

   • Bugs fixed: streaming, preserving reasoning deltas, guardrail timing, missing imports in examples, inert behavior in Anthropic model contexts etc.

     
     

7. MCP (Model Context Protocol) adoption & integration

   
   

   • OpenAI has adopted MCP (Model Context Protocol) across its products Agents SDK, Responses API) to standardize tool connectivity and context passing.

   • The Agent SDK supports deployment of MCP servers, and remote MCP servers are now usable in Realtime agents.

     
     

8. Backward compatibility & controls over breaking changes

   
   

   • The release process delineates how breaking changes will be handled (minor version increments vs patch).

   • There is guidance recommending pinning to certain major/minor versions if you don’t want surprises.